jerico
-
Posts
4 -
Joined
-
Last visited
Posts posted by jerico
-
-
Hello all, just want to share for Pascal-based GPU, Facial Recognition and DVA works if the platform is DVA3219 which has 1050Ti bundled.
I tried DVA3221 first. It "seemed" to work. GPU was detected in the info center, nvidia-smi shows running tasks. But when I tried Facial Recognition and Object Detection, nothing is actually being detected.
DVA3219 is available on arpl-i18n under "beta" platforms.
My setup is Proxmox + GTX 1060 3gb gpu passthrough.
- for Proxmox, had to blacklist Nvidia drivers in the host to be able to successfully passthrough.
-
My goal was to make NVENC work on Jellyfin.
Docker
I was able to expose my GPU in docker without the libnvidia-container by doing:
sudo docker run \ -e NVIDIA_VISIBLE_DEVICES=all \ -v /usr/local/bin/nvidia-smi:/usr/local/bin/nvidia-smi \ -v /usr/local/bin/nvidia-cuda-mps-control:/usr/local/bin/nvidia-cuda-mps-control \ -v /usr/local/bin/nvidia-persistenced:/usr/local/bin/nvidia-persistenced \ -v /usr/local/bin/nvidia-cuda-mps-server:/usr/local/bin/nvidia-cuda-mps-server \ -v /usr/local/bin/nvidia-debugdump:/usr/local/bin/nvidia-debugdump \ -v /usr/lib/libnvcuvid.so:/usr/lib/libnvcuvid.so \ -v /usr/lib/libnvidia-cfg.so:/usr/lib/libnvidia-cfg.so \ -v /usr/lib/libnvidia-compiler.so:/usr/lib/libnvidia-compiler.so \ -v /usr/lib/libnvidia-eglcore.so:/usr/lib/libnvidia-eglcore.so \ -v /usr/lib/libnvidia-encode.so:/usr/lib/libnvidia-encode.so \ -v /usr/lib/libnvidia-fatbinaryloader.so:/usr/lib/libnvidia-fatbinaryloader.so \ -v /usr/lib/libnvidia-fbc.so:/usr/lib/libnvidia-fbc.so \ -v /usr/lib/libnvidia-glcore.so:/usr/lib/libnvidia-glcore.so \ -v /usr/lib/libnvidia-glsi.so:/usr/lib/libnvidia-glsi.so \ -v /usr/lib/libnvidia-ifr.so:/usr/lib/libnvidia-ifr.so \ -v /usr/lib/libnvidia-ml.so.440.44:/usr/lib/libnvidia-ml.so \ -v /usr/lib/libnvidia-ml.so.440.44:/usr/lib/libnvidia-ml.so.1 \ -v /usr/lib/libnvidia-ml.so.440.44:/usr/lib/libnvidia-ml.so.440.44 \ -v /usr/lib/libnvidia-opencl.so:/usr/lib/libnvidia-opencl.so \ -v /usr/lib/libnvidia-ptxjitcompiler.so:/usr/lib/libnvidia-ptxjitcompiler.so \ -v /usr/lib/libnvidia-tls.so:/usr/lib/libnvidia-tls.so \ -v /usr/lib/libicuuc.so:/usr/lib/libicuuc.so \ -v /usr/lib/libcuda.so:/usr/lib/libcuda.so \ -v /usr/lib/libcuda.so.1:/usr/lib/libcuda.so.1 \ -v /usr/lib/libicudata.so:/usr/lib/libicudata.so \ --device /dev/nvidia0:/dev/nvidia0 \ --device /dev/nvidiactl:/dev/nvidiactl \ --device /dev/nvidia-uvm:/dev/nvidia-uvm \ --device /dev/nvidia-uvm-tools:/dev/nvidia-uvm-tools \ nvidia/cuda:11.0.3-runtime nvidia-smiOutput is:
> nvidia/cuda:11.0.3-runtime nvidia-smi Tue Aug 1 00:54:12 2023 +-----------------------------------------------------------------------------+ | NVIDIA-SMI 440.44 Driver Version: 440.44 CUDA Version: N/A | |-------------------------------+----------------------+----------------------+ | GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC | | Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. | |===============================+======================+======================| | 0 GeForce GTX 106... On | 00000000:01:00.0 Off | N/A | | 84% 89C P2 58W / 180W | 1960MiB / 3018MiB | 90% Default | +-------------------------------+----------------------+----------------------+ +-----------------------------------------------------------------------------+ | Processes: GPU Memory | | GPU PID Type Process name Usage | |=============================================================================| +-----------------------------------------------------------------------------+This should work on any platform that has NVIDIA runtime library installed.
However, this still does not seem to work with Jellyfin docker. I can configure NVENC, play videos fine, but the logs does not show h264_nvenc. I also see no process running in `nvidia-smi`.
Official docs points to using nvidia-container-toolkit https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/install-guide.html
That's why I was looking at how to build it with DSM 7.2 kernel.
Running rffmpeg
My second idea was to use rffmpeg (remote ffmpeg to offload transcoding to another machine). I was thinking running Jellyfin in Docker and configure rffmpeg, then run the hardware accelerated ffmpeg in DSM host.
I downloaded the portable linux jellyfin-ffmpeg distribution https://github.com/jellyfin/jellyfin-ffmpeg/releases/tag/v5.1.3-2
Running it in ssh yields
[h264_nvenc @ 0x55ce40d8c480] Driver does not support the required nvenc API version. Required: 12.0 Found: 9.1 [h264_nvenc @ 0x55ce40d8c480] The minimum required Nvidia driver for nvenc is (unknown) or newer Error initializing output stream 0:0 -- Error while opening encoder for output stream #0:0 - maybe incorrect parameters such as bit_rate, rate, width or height Conversion failed!
I think this is because of the driver DSM uses which is an old 440.44. jellyfin-ffmpeg is compiled with the latest https://github.com/FFmpeg/nv-codec-headers. The DSM Nvidia driver only supports 9.1.
Confirming NVENC works
I confirmed NVENC works with the official driver by installing Emby and trying out their packaged ffmpeg
/volume1/@appstore/EmbyServer/bin/emby-ffmpeg -i /volume1/downloads/test.mkv -c:v h264_nvenc -b:v 1000k -c:a copy /volume1/downloads/test_nvenc.mp4
+-----------------------------------------------------------------------------+ | Processes: GPU Memory | | GPU PID Type Process name Usage | |=============================================================================| | 0 15282 C /var/packages/EmbyServer/target/bin/ffmpeg 112MiB | | 0 20921 C ...ceStation/target/synoface/bin/synofaced 1108MiB | | 0 32722 C ...anceStation/target/synodva/bin/synodvad 834MiB | +-----------------------------------------------------------------------------+
Next steps
The other thing I have yet to try is recompile jellyfin-ffmpeg with an older nv-codec-headers and use it inside Jellyfin docker
-
On 11/30/2021 at 8:31 PM, Zowlverein said:
Have got cuda in docker working (at least nvidia-smi with the sample container: sudo docker run --gpus all nvidia/cuda:10.2-runtime nvidia-smi)
Required files here: https://gofile.io/d/zYTBCP
/usr/bin/nvidia-container-toolkit (v 1.5.1 from ubuntu 16.04 build)
/usr/bin/nvidia-container-cli (custom build with syscall security removed)
also need symlink (ln -s /usr/bin/nvidia-container-toolkit /usr/bin/nvidia-container-runtime-hook)
/lib/libseccomp.so.2.5.1 (from ubuntu 16.04 - also ln -s /lib/libseccomp.so.2.5.1 /lib/libseccomp.so.2)
/lib/libnvidia-container.so.1.5.1 (from same custom build - also ln -s /lib/libnvidia-container.so.1.5.1 /lib/libnvidia-container.so.1)
ldconfig - copied to /opt/bin/ from ubuntu installation (version in entware/optware didn't work)
/etc/nvidia-container-runtime/config.toml (need to update file to point to local ldconfig and nvidia drive paths)
I know this is an old post, but can anybody provide how to build the `libnvidia-container` for DSM 7.2 kernel?
DVA3219 loader development thread
in Developer Discussion Room
Posted · Edited by jerico
I have DVA3219 working with GT 1030 with Facial Recognition and Deep Video Analytics. I'd say a good low-power alternative to enable AI tasks.
I used arpl-i18n loader. DVA3219 is under beta platform.
+-----------------------------------------------------------------------------+ | NVIDIA-SMI 440.44 Driver Version: 440.44 CUDA Version: 10.2 | |-------------------------------+----------------------+----------------------+ | GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC | | Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. | |===============================+======================+======================| | 0 GeForce GT 1030 On | 00000000:01:00.0 Off | N/A | | 58% 60C P0 N/A / 30W | 1516MiB / 2001MiB | 74% Default | +-------------------------------+----------------------+----------------------+ +-----------------------------------------------------------------------------+ | Processes: GPU Memory | | GPU PID Type Process name Usage | |=============================================================================| | 0 26525 C ...ceStation/target/synoface/bin/synofaced 746MiB | | 0 26527 C ...anceStation/target/synodva/bin/synodvad 760MiB | +-----------------------------------------------------------------------------+